A new Application of “Gemination” in Standard Arabic Language
- Application for the voice synthesis -
OUARDA HACHOUR AND Nikos Mastorakis
Development Scientific Center of Advanced Technologies and Technical research
For the Development of the Arabic Language (C.R.S.T.D.L.A)
1,rue Djamel Eddine al-Afghani –Bouzareah
Algiers Algeria
Phone/fax : (213) (021) 94-12-38
Hellenic Naval Academy
Terma Hatzikyriakou, 18539
Piraeus, Greece
Abstract: - This work contributes to solve the problem of the”tachdid” in Standard Arabic language, in order to synthesize speech using the technical of Mbrola and to reduce the VC1V (Vowels, Consonant, Vowels) units of the dictionary in a speech synthesis. The number of pre-stored units can be diminished to the half thanks to the modelisation of the V C2V, where C2 represents the geminate consonant C1. This maybe possible by a modification of some acoustical parameters of the corresponding wave's form of the VCV units. We have judged useful to study the modifications undergone by the curve of the energy, the temporal durations and the formantic values (features) of the subsequent vowels using the Computering Speech Laboratory CSL of Kay Elemetrics and the Praat and Matlab Windows Programs. For the application, we have utilized the technical Mbrola (Multi Band Re-synthesis Overlap Add) witch is a most “efficacy” technique to have the natural speech.
Keywords: synthesis, MBrola, Voice synthesis, gemination, natural speech, duration.
1 Introduction
The " tachdid ", indicated in Standard Arabic Language by the sign ( ّ ) called cheddah, is often confused wrongly or rightlywith the French term " gemination ". This characteristic isone of the most significant phenomena characterizing the Standard Arabiclanguage. All the consonants of this language can be geminated exceptfor the glottalcharacter [ Hamza ].
The theory and practice of this concept are currently among the most intensively studied and promising areas in computer science and engineering which will certainly play a primary goal role in future. These theories and applications provide a source linking all fields in which intelligent control plays a dominant role.
All research areas in this domain showthe variations on which the bodies of the vocal apparatus undergo at the
time of the resulting of gemination and acoustic effects. It remainsthat the majority of its work do not converge
the sameresults. Controversies sullied old work, in particular thosereferring to the modifications that undergo the movements of thebodies of the vocal apparatus and the acoustic effects resulting.This phenomenon which presents only one simple means of expression inmuch of languages like French, is very relevant in Arabic.
In this paper we studied clearing up some acoustic effects,resulting from the pronunciation of a consonant geminated inopposition to its corresponding not geminated in the same context andits possible application in voice synthesis. We studied themodifications which undergo the temporal durations and the formanticvalues of the subsequent vowels. The shape of the curve of energyduring the pronunciation of a phoneme was compared with its correspondent geminated in the same context. The presentation is organized as follow: In section II, the proposed The “tachdid” in Arabic language is done. Section III describes The Acoustic study, Section IV presents The Application of “Tachdid ” in Standard Arabic Language:
and implementation results and finally, a conclusion is given in Section V.
II The “tachdid”in Arabic language
Various definitions were givento the concept of Standard Arabic tachdi, This concept was very old, but it is the most research area based on the concept of Sibawayh, quoted by A. Roman in its studyof Arab phonology. For Sibawayh, " it is heavy to employ its languagewith leaving a place of articulation for making there at once return. Also, because of this tiredness which the realization of twoidentical articulations brings, this realization is ejected infavour of the gemination of two [ harf ] identical ones so thatthere is only one rise in the language " [ 1 ].
In a general way, the tachdid is defined as phenomenon of reinforcement of the consonant articulation which prolongs the duration of it approximately half and the intensity increases some. This phenomenon is sometimes called dedoublement, although there is not truly repetition of the consonant. A consonant can be also comparable in the pronunciation by a different consonant which follows it, causing a gemination because of a heaviness articulator.
In this area, previous research on language synthesis using a robust tool to build a good synthesis. The major thrust of this type of tools is to obtained a really synthesis and to test the capability of a system in order to offer a natural speech and giving a real intelligent component, short time of execution, high speed processing and the objectives of all speech synthesis are obtained. We can denote, the treatment used MEL-frequency Cepstrum Coefficient (MFCCs) extensively for parameterisation of speech and reasonably good results were obtained on varied classes of language synthesis [1]. In [1], the results were that the Linear Prediction Cepstrum Coefficients ( LPCC) out perform med-frequency Cepstrum Coefficients ( MFCC) in all tests. Also, one commonly used technique for synthesis of the speech wave form in text-to-speech synthesis is concatenation of short speech units taken from a pre-recorded in ventory [2] .After concatenation , these units are modified in duration and “melody” to smoothly join each other and achieve the prosody of a natural utterance. These modifications can be performed without introducing unnatural-sounding arti-facts; signal modeling techniques, such as the popular PSOLA technique must be employed. Sinusoidal signal models have been shown to be useful for speech prosody modification and speech synthesis [3].
In the context of germination, and according to a studymade by M.C. Dkhissi-Boff, " the geminated consonant does not presenttwo distinct articulator movements, but only one single movement,which differs from that of the simple consonant, by its greatarticulator stability and its very significant duration " [ 4 ]. ForP. Delattre, the articulation of geminated is carried out in twophases and presents two tops of activity. On the other hand, forRousselot, J Cantineau, J.F Bonnot and M.C. Dkhissi-Boff, thegeminated consonants are carried out in only one phase of greatarticulator stability in which, the duration is definitely affected. Other work explains that one duration of the preceding vowel inverselyproportional to the articulator force of the subsequent consonant [6].
III Acoustic study
We used for theacoustic study, module CSL (Computer Speech Laboratory) for Windows, Kay Elemetrics, and the software of analysis and treatment:Matlab andPraat.
The study was carried out starting from several records which are different. We practically raised same theconstatations with the same locution.We have recorded thevalues of the formants of the vowels following (V) at the same time of thetransitions, as well as the temporal durations (D) (in milliseconds)of the preceding vowel (PV), and the Next vowel (NV), geminatedconsonant (c1) and sound opposed not geminated (c2).
III.1 The duration Study
The average durations of the geminated / not geminated consonants areillustrated in the table 1.
We calculate D1, D2, D3 which are presented by :
D1 = PV2 / PV1, where PV2 and PV1 are the duration of vowels which precedes C1et C2; respectively.
D2 = NV2 / NV1, where NV2 et NV1 are the duration of posterior vowels of C1 and C 2; respectively.
D3 = Ds1 / Ds2, where Ds1and Ds2 are the duration of the consonants C1 and C2; respectively. (See table 2)
DurationPhoneme / PV / NV / D
[t] / 80.2 / 92.5 / 109.6
[tt] / 65.5 / 109.3 / 229.6
[d] / 71.5 / 64.2 / 94.1
[dd] / 71.4 / 76.1 / 200.6
[b] / 65.3 / 73.8 / 72.8
[bb] / 51.8 / 84.5 / 209.8
[y] / 65.4 / 71.1 / 112.4
[yy] / 51.0 / 76.2 / 212.2
[k] / 76.9 / 111.3 / 104.8
[kk] / 57.8 / 126.4 / 201.8
[‡] / 72.5 / 99.4 / 63.0
[‡‡] / 45.8 / 100.5 / 178.4
[x] / 81.0 / 93.6 / 108.2
[xx] / 66.2 / 101.4 / 195.3
Table 1: The average duration s of the geminated / not geminated consonants
By analyzing these results, we have notedthe following remarks:
-The durations of the geminated consonantare significant when it is compared toits opposite notgeminated;
- The larger duration is the duration of the vowel whichfollows the onegeminated.
-A fall durationis the duration of the precedingvowel in the presence of one geminated.
In the Figure 1; we presents the curves of energy during the pronunciation of phonemesocclusive [ T. ] and fricative [ H ] compared with those theirgeminated correspondents.The study of the various curves of energy shows us that:
- The duration of occlusion of the geminated consonant is much moresignificant than its opposite not geminated
- The curve ofamplitude of energy continues to go down because of lengthening fromthe duration from the behaviour from the occlusive consonant. We can
Nevertheless note that the spectrum of the geminated consonantpresented, in practically all the studied cases, energy uniformlyleft again and without notable discontinuity. In this direction, itis difficult to separate the end from the first consonant and thebeginning of the second consonant.
That converges towards thedirection of the thesis of the realization of the consonant geminatedin only one phase of greatarticulator stability.
In the case of consonants not geminated, we presents the following values in the case of given the assimilation: [ hal yaqûl ]: lasted of emission of [ y ] the0.065s approximately, value of the formants of the vowel whichfollows to it [ y ] F1520Hz, F22240Hz and F32640Hz. [ hayyaqûl ]: the lasted emission of the geminated consonant [ yy ] = 0.150s, valueof the formants of the vowel which follows geminated [ yy ]F1440Hz, F22200Hz and F32680Hz.
III.2 Results
The results obtained show that at the time of thegemination, it generally occurs(See Figure 2 , and Figure 3):
-A light fall of the formants F1and F2 of the vowel which follows the consonant geminated compared tothe not geminated consonant.
-The value of the formant F1 of the preceding vowel also undergoes a fall. As regards the formant F3, onenotes in much case a light rise in the presence of the geminated consonant.
-The duration of occlusion of the geminated consonantis very important.
-A fall of the curve of rather significant energyduring the behaviour of the occlusion.
-The largest duration is when thefollowing vowel is geminated.
-A reduce inthe duration is produced when the preceding vowel is in the presence of geminated one.
Coefficientsphonèmes / D1 / D2 / D3
[x] / 0.83 / 1.11 / 1.79
[b] / 0.68 / 1.19 / 2.79
[k] / 0.77 / 1.09 / 1.06
[l] / 0.61 / 1.19 / 2.68
[t] / 0.80 / 1.12 / 2.02
[d] / 0.90 / 1.09 / 2.09
assimilation / 0.82 / 1.03 / 1.73
Table 2: Coefficients of the reports / ratios of the durations
Figure2. Spectrographic representations of the sentence : (alama eddarsa)
Figure3. Spectrographic representations of the sentence: (allama eddarsa)
Figure 3: curves of energy during the pronunciation of occlusive phonemes [t] compared at its opposite geminated
Figure 3: curves of energy during the pronunciation of fricative phonemes [h] compared at its opposite geminated
IV.The Applicationof “Tachdid ” in
Standard Arabic Language:
The application is done using the technique .MBROLA 5.1 (Multi Band Re-synthesis Overlap Add) is atechnique of voice synthesis founded on the concatenation of thesounds. The principal advantage of a synthesis of word atMBROLA [7] base is that it permits:
- An appreciably reduced memory size of the dictionary
of dip hones.
- A word of synthesis is very fluid.
- The smoothing of spectraldiscontinuities appearing of share and others of the points ofconcatenation
-The amount of degradation introduced by this modification is highly dependent on the actual f0 curve of the original recordings. Non-uniform unit based synthesis aimed speech corpora are far away from having constant f0 since a good coverage of general prosodic features of natural speech is needed.
-To improve the quality o f synthesis, database pre-processingin the NUMBROLA project has recently been modified such that the harmonic re-synthesizer re-synthesizes frames with their original pitch and phase envelope. This change improves the quality o f copy-synthesis to a degree very close to transparency.
For ourApplication,we have used a sound data dictionary of 1650 unitsof variable sizes (polysounds). For the huge memory of data stored, we havecompressed this database for its exploitation starting from the waiter of the application. It is possible to modify the words by better an improvement of the natural speech, modifying the relevant parameters which are the durationand the pitch (by respecting the models obtained during the acousticanalysis). The quality of the word obtained is rather good.
V Conclusion
This experiment showed us that the contribution ofthe MBROLA technique in the voice synthesis in Standard ArabicLanguage is very interesting especially when we re-synthesis the word. Theinsertion of the MBROLA module, in the flow offers a very profitable sight quality of the synthesis word obtained. Moreover this work showed us that by anadequate modelling of the phonemes geminated in the various contexts,we can obtain a synthetic word without having to resort to the designof the sound units geminated in the dictionary of sound units of the database. That will enable us to reduce the size of it considerably.This reduction is of as much significant.
References:
[1] E.Wong and S.Sridharan , Comparison of linear prediction Cepstrum coefficients and Mel-Frequency Cepstrum Coefficients for Language Identification, School of Electrical and electronic Systems Engineering, Queensland University of Technology.
[2] M.W.Macon and M.A Clements, Speech concatenation and synthesis using an overlap-add sinusoidal model, School of Electrical and Computer Engineering Georgia Institute of technology , Atlanata , GA30332-0250.
[3] E. Moulinss and E.Charpenter, “Pitch synchronous a wave processing techniques for text-to-speech synthesis using dip hones”, Speech Communication , vol.9, pp453-467, December 1990.
[4] M.C. Dkhissi-Boff . Contribution à l'étude expérimentale des consonnes de l'arabe. Travaux de l'Institut de Phonétique de Strasbourg, N 15, 1983;
[5] - D. Ozza Obaid . Fan attadjouid. Dar Ibn Hazm, Beyrout, Liban, 1991 .
[6] - M. Attaoui. Force articulatoire et gémination en arabe marocain de Fès. Travaux de l'Institut de Phonétique de Strasbourg, T.I.P.S. 23, 1993, Page 22, 1993;
[7] B. Bozkurt, T. Dutoit,R. Prudon, C. D’Alessandro, V. Pagel, and M. ASBL, IMPROVING QUALITY OF MBROLA SYNTHESIS FOR NON-UNIFORM UNITS SYNTHESIS, Initialis Sci.Park,B-7000 Mons,Belgium LIMSI, CNRS,Po Box 133 –F91403 Orsay,France.